Evaluation library

This is the fastest way to evaluate a model. The library provides a C API and a simple C++ wrapper API. The C API interface can be accessed from any programming language.

Download

Prebuilt shared library artifacts are available.

An up-to-date list of available CatBoost releases and the corresponding binaries for different operating systems is available in the Download section of the releases page on GitHub.

Operating system CPU architectures GPU support using CUDA
macOS (versions currently supported by Apple) x86_64 and arm64 no
Linux (compatible with manylinux2014 platform tag ) x86_64 and aarch64 yes
Windows 10 and 11 x86_64 yes

Note

Release binaries for x86_64 CPU architectures are built with SIMD extensions SSE2, SSE3, SSSE3, SSE4 enabled. If you need to run CatBoost on older CPUs that do not support these instruction sets build CatBoost artifacts yourself

Operating system CPU architectures GPU support using CUDA
Linux (compatible with manylinux2014 platform tag ) x86_64 and aarch64 yes
macOS (versions currently supported by Apple) x86_64 and arm64 no
Windows 10 and 11 x86_64 yes
Operating system Files
Linux libcatboostmodel-linux-{cpu_arch}-{release_version}.so
macOS libcatboostmodel-darwin-universal2-{release_version}.dylib
Windows catboostmodel-windows-{cpu_arch}-{release_version}.lib and catboostmodel-windows-{cpu_arch}-{release_version}.dll

Applying on GPU requires NVIDIA Driver of version 450.80.02 or higher.

Note

Only CUDA 11.0 is officially supported in compiled binaries for Windows. Вuild the binary from a local copy if GPU support is required and the installed version of CUDA differs from 11.0. CatBoost should work fine with CUDA 11.0 and later versions.

All necessary CUDA libraries are statically linked to the Linux and macOS binaries of the CatBoost applier library, therefore, the only installation necessary is the appropriate version of the CUDA driver.

Note

Release binaries for x86_64 CPU architectures are built with SIMD extensions SSE2, SSE3, SSSE3, SSE4 enabled. If you need to run CatBoost on older CPUs that do not support these instruction sets build library binary artifacts yourself

Build from source

Warning

CatBoost uses CMake-based build process since this commit. Previously Ya Make (Yandex's build system) had been used.

Select the appropriate build method below accordingly.

Source code

CatBoost source code is stored as a Git repository on GitHub at https://github.com/catboost/catboost/. You can obtain a local copy of this Git repository by running the following command from a command line interpreter (you need to have Git command line tools installed):

git clone https://github.com/catboost/catboost.git

Build using CMake

Build catboostmodel target.

See Build native artifacts.

Built artifacts will be in $CMAKE_BINARY_DIR/catboost/libs/model_interface:

OS Files
Linux libcatboostmodel.so
macOS libcatboostmodel.dylib
Windows catboostmodel.lib and catboostmodel.dll

Build catboostmodel_static target.

See Build native artifacts.

Built library will consist of two parts:
- global part. This part contains symbols that require forced initialization.
- non-global part. All other symbols.

Built artifacts will be in $CMAKE_BINARY_DIR/catboost/libs/model_interface/static:

OS Files
Linux or macOS libcatboostmodel_static.a, libcatboostmodel_static.global.a
Windows catboostmodel_static.lib, catboostmodel_static.global.lib

Build using Ya Make

  1. Setup build environment

  2. Open the catboost directory from the local copy of the CatBoost repository.

  3. Run the following command:

    ./ya make -r [optional parameters] catboost/libs/model_interface
    

    The output directory catboost/libs/model_interface will contain:

    OS Files
    Linux libcatboostmodel.so
    macOS libcatboostmodel.dylib
    Windows catboostmodel.lib and catboostmodel.dll
    ./ya make -r [optional parameters] catboost/libs/model_interface/static
    

    The output directory catboost/libs/model_interface/static will contain a pair of artifacts:

    • liblibcatboostmodel.o. This part contains symbols that require forced initialization.
    • libcatboostmodel.a. This part contains all other symbols.

    Useful parameters:

    Parameter Description
    -DCUDA_ROOT The path to CUDA. This parameter is required to support training on GPU.
    -DHAVE_CUDA=no Disable CUDA support. This speeds up compilation.

    By default, the package is built with CUDA support if CUDA Toolkit is installed.

Build using Make (Linux-only)

Warning

This approach will work only for versions prior to this commit.

For newer versions use Build with CMake

Choose the preferred way to use the evaluation library and compile it accordingly:

export CXX=/path/to/clang++
export CC=/path/to/clang

make -f make/model_interface.CLANG50-LINUX-X86_64.makefile

The output directory catboost/libs/model_interface will contain libcatboostmodel.so.

export CXX=/path/to/clang++
export CC=/path/to/clang

make -f make/model_interface_static.CLANG50-LINUX-X86_64.makefile

The output directory catboost/libs/model_interface/static will contain a pair of artifacts:

  • liblibcatboostmodel.o. This part contains symbols that require forced initialization.
  • libcatboostmodel.a. This part contains all other symbols.

Usage

The CatBoost model can be loaded from a file or initialized from the buffer memory.

C API

Perform the following steps to use this API:

  1. Use the methods from the c_api.h file (refer to the doxygen-style documentation for details).

    Sample C code without include statements:

    float floatFeatures[100];
    char* catFeatures[2] = {"1", "2"};
    double result[1];
    ModelCalcerHandle modelHandle;
    modelHandle = ModelCalcerCreate();
    if (!LoadFullModelFromFile(modelHandle, "model.cbm")) {
        printf("LoadFullModelFromFile error message: %s\n", GetErrorString());
    }
    if (!CalcModelPrediction(
            modelHandle,
            1,
            &floatFeatures, 100,
            &catFeatures, 2,
            &result, 1
        )) {
        printf("CalcModelPrediction error message: %s\n", GetErrorString());
    }
    ModelCalcerDelete(modelHandle);
    
  2. Add the required libraries to the linking command.

    Linker is often invoked through the compiler call, examples below assume that.

    • Linux or macOS

      Example:

      clang++ <your sources and options> -L<path_to_dir_with_libcatboostmodel> -lcatboostmodel
      
    • Windows

      Example:

      cl.exe <your sources and options> /link <path_to_dir_with_libcatboostmodel>\catboostmodel.lib
      

    The shared library must be accessible from the dynamic library loader search path. See your operating system documentation for the details.

    Add both global and non-global parts to the linker input. global part requires passing special platform-specific flags to force the required initialization of symbols.

    See per-platform examples below:

    • Linux

      On Linux additional libraries libdl and libpthread have to be added to the linker input as well.

      clang++ <your sources and options> -nodefaultlibs -lpthread -ldl -Wl,--whole-archive <catboost_lib_dir>/libcatboostmodel_static.global.a -Wl,--no-whole-archive <catboost_lib_dir>/libcatboostmodel_static.a
      
    • macOS

      clang++ <your sources and options> <catboost_lib_dir>/libcatboostmodel_static.a
      -Wl,-force_load,<catboost_lib_dir>/libcatboostmodel_static.global.a
      
    • Windows

      When using c_api.h with the static library the additional define CATBOOST_API_STATIC_LIB is required.

      cl.exe <your sources and options> /DCATBOOST_API_STATIC_LIB /link /WHOLEARCHIVE:<catboost_lib_dir>\catboostmodel_static.global.lib <catboost_lib_dir>\catboostmodel_static.lib
      

    Add both liblibcatboostmodel.o and libcatboostmodel.a to the linker input.

    On Linux additional libraries libdl and libpthread have to be added to the linker input as well.

    Example:

    clang++ <your sources and options> liblibcatboostmodel.o libcatboostmodel.a -ldl -lpthread
    

C++ wrapper API

A C++ wrapper for the C API interface is also available.

Refer to the wrapped_calcer.h file and the sample CMake project in the CatBoost repository for more details.

Usage example:

ModelCalcerWrapper calcer("model.cbm");
std::vector<float> floatFeatures(100);
std::vector<std::string> catFeatures = {"one", "two", "three"};
std::cout << calcer.Calc(floatFeatures, catFeatures) << std::endl;

ModelCalcerWrapper also has a constructor to read data from the memory buffer.

Source code and a CMake usage example